Authored compression

Written once, by the person who knows

A README compresses a folder. A header compresses a file, a case line compresses a test, a log compresses an agent’s run, and a ref compresses a hand-off. Each is written by whoever knew the most about that part, when they knew it best.

Nothing has to summarise it later. The compression is done once, when it is written, and every reader after gets the result.

Three levels to read it

memex reads the compression back one level at a time, and an agent should too: the one-liner first, then the document, then the chain above it.

• memex search prints the one-liners of what matches, across the whole layer.
• memex show prints one document, with what it cites and what cites it.
• The chain is the READMEs above that document, so its context comes with it.

$ memex show crates/namer               # README, children, chain
$ memex show crates/namer/src/hash.rs   # header, cites, cited by, chain

Read the one-liner before the document, and the document before the code. Open code only when the layer cannot answer.

Why there is no server, database or model

The summaries are already written, so memex only has to read them. It is Markdown and a CLI: it starts, reads the plain files as they are, prints and exits.

So it needs no server, daemon, database, model or vectors. Each of those would redo a compression that the author has already done.

The trade-off is that a summary is only as current as its last edit. That is what drift is about.

See Context compression on the intro page, and the verbs for their flags.